FL-SEC: Privacy-Preserving Decentralized Federated Learning Using SignSGD for the Internet of Artificially Intelligent Things
Youyang Qu, Chenhao Xu, Longxiang Gao, Yong Xiang, Shui Yu
Abstract
The wide proliferation of the Internet of Things (IoT) and the unimaginable rapid advance of artificial intelligence (AI) jointly facilitate the Internet of Artificially Intelligent Things (A-IoT). Artificially Intelligent things (AI-T) run machine learning (ML) models locally while interacting and exchanging data with other AI-Ts through A-IoT. However, sensitive data may be abused during transmission by malicious or compromised AI-Ts. Federated learning is thereby proposed to achieve secure communication, where AI-Ts maintain the same ML model by exchanging model parameters instead of raw data. However, there are three significant issues for FL being applied in A-IoT. First, the context of model parameters of each AI-T may leak privacy, resulting from inference attacks. Second, falsified-data-based poisoning attacks may lead to a failure of ML model convergence. Third, exchanging model parameters cost more communication resources than expected in this resource-limited scenario. To address these issues, we propose privacy-preserving decentralized FL for secure and efficient communication (FL-SEC) over A-IoT. A novel blockchain structure is devised to achieve decentralized FL, which prevents single point of failure and poisoning attacks and also provides a personalized incentive mechanism. In addition, improved sign gradient descent is used to replace traditional gradient descent, which preserves the privacy of model parameters while significantly reducing the communication resource consumption. Experiments on real-world benchmark datasets show the superior performance of the proposed model.